acceptodds
Under review as a conference paper at ICLR 2027

Topological Path Signatures under Physical Micro-Fractures: Pareto-Optimal Battery Degradation Estimation on Edge Microcontrollers

Abstract

Sequence learning paradigms demonstrate low empirical error on curated laboratory battery benchmarks, yet rely upon an unstated operational prior: continuous, uniformly sampled charging trajectories globally aligned to a standardized temporal axis. In commercial vehicular deployment, this prior is violated by physical micro-fractures, including unanchored opportunistic charging slices (- State-of-Charge intervals), sensor quantization floors ( ADC steps), and stochastic communication dropouts. Under this manifold, continuous spline-based ODE solvers encounter numerical stiffness deadlocks, self-attention mechanisms undergo key-padding dilution, and unrolled hidden states exceed edge microcontroller (MCU) memory bounds. In this work, we present RIIS-MicroBurst, an edge-native framework utilizing a parameter-free feature extraction operator formulated in free Lie algebra. By replacing numerical integration with deterministic tensor-vector operations via Chen's identity, RIIS maps arbitrary variable-length inputs into a static, invariant geometric representation. Combined with thermodynamic phase-plane scaling and a closed-form Geometry-Oracle, RIIS establishes a new Pareto efficiency frontier across simulation, dynamic stress, synthetic fault, and commercial fleet benchmarks. Operating strictly within automotive MCU hardware ceilings ( SRAM, FLOPs), it achieves state-of-the-art accuracy while delivering to lower computational operations compared to recurrent and continuous baselines.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Machine Learning, Path Signatures, Deterministic Routing, TinyML, Resource-Constrained AI, Battery Degradation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jfCjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RIIS-MicroBurst, a lightweight framework for battery degradation estimation from short, irregular, quantized, and partially missing telemetry. It combines truncated log-signatures with physically motivated feature weighting, degree-wise normalization, and geometry-based model selection. Experiments across simulated, laboratory, synthetic fault, and fleet datasets report competitive predictive performance and substantial computational savings for microcontroller deployment.

Critical strengths
  1. Practical problem relevance: The paper addresses battery monitoring under irregular telemetry and computational constraints.
  2. Potential for efficient inference: Incrementally computable, fixed-dimensional path features offer a plausible approach to reducing inference memory requirements.
  3. Evidence across multiple datasets: The reported results suggest potential benefits for degradation estimation and fault classification across several settings.
Critical weaknesses
  1. Incomplete efficiency accounting: The current analysis does not clearly establish whether all feature extraction costs and optimized streaming baselines are included, leaving the extent of the deployment advantage uncertain.
  2. Robustness claims need further support: Reparameterization invariance alone does not establish robustness to missing samples or quantization, as these can alter path geometry.
  3. Reproducibility requires clarification: Differences between the method description and reference implementation, together with missing evaluation details, make the reported results probably difficult to independently assess.
Questions for authors
  1. End-to-end efficiency: Could the authors clarify whether the computational advantages persist when feature extraction costs and optimized streaming baselines are included?
  2. Implementation consistency: Could the authors clarify how the provided implementation corresponds to the models used in the reported experiments?
  3. Robustness assumptions: Could the authors specify the assumptions underlying robustness to missing samples and quantization, and provide supporting evidence?
  4. Evaluation protocol: Could the authors clarify the data partitioning and prediction targets, particularly how they support independent evaluation of battery health estimation?
Reviewer_SyeeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RIIS-MicroBurst studies battery-related prediction from short, irregular telemetry sequences under tight edge-computing budgets. It represents voltage, current, temperature, and sampling intervals as a multichannel path. Truncated log-signatures map this path into a fixed-dimensional representation through algebraic accumulation. The framework reweights selected second-order coordinates to emphasize voltage–current and voltage–temperature interactions. A degree-wise normalization layer rescales different signature orders before prediction. A geometry-based routing rule selects the signature depth and regularized readout, while an MLP variant provides a more expressive prediction head. The intended contribution is a compact alternative to sequence models for quantized and partially observed charging trajectories.

Critical strengths
  1. The paper addresses a relevant deployment problem involving short charging segments, missing telemetry, and constrained hardware. These conditions deserve more attention than evaluations restricted to complete laboratory cycles.
  2. Fixed-dimensional path features combined with lightweight readouts are a sensible direction for resource-constrained prediction. Algebraic accumulation also provides a natural mechanism for processing variable-length inputs.
  3. The manuscript includes component ablations and comparisons between intact and corrupted trajectories. These are relevant tests for examining the contributions of the representation, physical weighting, and routing choices.
Critical weaknesses
  1. Cropping, quantization, and packet deletion change the geometric path rather than merely its parameterization. Reparameterization invariance therefore does not establish the claimed immunity to these corruptions, especially when sampling intervals are themselves an input coordinate.
  2. The supplied FLOP audit counts only linear and normalization layers, omitting the sequence-dependent signature extraction. Together with workstation-based latency profiling, this leaves the headline speedups and MCU feasibility without a consistent end-to-end validation.
  3. Several claimed baseline failures are implementation-dependent rather than structural. An LSTM does not require retaining an entire sequence of activations during inference, and correctly masked attention does not dilute weights simply because padded keys are present.
  4. The supplied degree-4 extractor computes only the first ten features and fills the remaining eighty coordinates with zeros. The actual higher-order implementation is deferred to release, so the provided code cannot reproduce the central degree-4 results.
  5. The paper does not specify cell- or vehicle-disjoint train/validation/test splits or explain how related charging slices are separated. The IVST target is also described as “odometer degradation,” leaving its connection to SOH or capacity estimation unclear.
Questions for authors

See Weakness.

Reviewer_7XfoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses 4 key problems in state-of-charge estimation for batteries of electrical vehicles. The first three problems are related to domain shift between model training time and deployment, the fourth is related to computational and storage cost of the model in-field. More in details: the assumption of continuous physical models is is violated by physical micro-fractures and opportunistic partial charging, the prior of high percision sensor readings is violated by high real-life sensor quantization floors, finally the assumption of regular readings from sensors is violated by communication drop-outs. The paper addresses these "domain shift" challenges with a novel non-parametric feature extraction operator. Furthermore, the challenge of limited computational and storage in the hardware used for model deployment is also addressed in part thanks to the feature extraction operators leading to models that are substantially smaller and less computationally intensive than the SoA, while achieving competitive accuracy, thereby defining a new Pareto frontier.

Critical strengths
  1. solid mathematical grounding of the method proposed, as opposed to pure data-centric approaches. The mathematical formulation is hard to follow for the non expert, but its link to the key non-idealities of the battery discharge in field is clear and intuitive.
  2. Significant speedup and reduction of model footprint, while at the same time achieving competitive accuracy, superior to several methods that require orders of magnitude more resources.
Critical weaknesses
  1. the "non idealities" (low resolution of the measurements, physical fractures in the materials, incomplete opportunistic recharge cycles are states multiple times but there is no compelling evidence (from literature or new experiments) that these are general problems for all embedded battery recharging scenarios, or that they are the main problems. For instance, it is widely known that charge delivery and effective state of charge is extremely temperature dependent - also in term of temperature cycles during charge and discharge. This aspect is known to dominate in the practice of electrical vehicle usage and it is not discussed in this papers. Similarly, 10mV is a pretty modest accuracy as many A2D have 20 bit of resolutions. So it is not clear how general is the assumption Finally, also the packet loss seems to be quite arbitrary. I do not doubt that there are cases where the model is accurate, but when e.g. one uses reliable field busses, the losses could be minimal
  2. While the "effort estimation" for embedded implementation is reasonable (flops and bytes), it is known to be a quite lose proxy for deployability in embedded systems - and this lack of real implementation on an MCU is weakening the strength of the final claims.
Questions for authors
  1. Discuss temperature effects
  2. provide more evidence of the usability of the methods even in case when one or more of the assumed non idealities is not present or has a different impact (e.g. reliable communication, or high resolution A2Ds)
  3. provide more evidence of implementability on a real target (e.g. a commercial ARM Cortex MCU)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.